Graph Embedding Structure Data Analysis Technique And Applications

The graph embedding structure data analysis technique is studied in this thesisin order to solve a common issue: the visualization of the high dimensions andmultiple elements dataset. This issue also belongs to one of the critical challengeissues in the area of intelligent test data analysis of spacecraft and aircraft, whichplays an important roll in the fault diagnose and prognosis of system. The graphembedding data analysis techniques can preserve the essential structure of datasetwhile feature mapping. The essential feature can be presented visually if the highdimension dataset are maped into low dimension. Analysing the advantages anddisadvantages of the existing techniques, new graph embedding structure dataanalysis techniques are studied. Taking example by the locality preservingprojections (LPP), improved algorithms are studied and used in the visualizationanalysis of satellite remote sensing data.Firstly, as for LPP only considers the locality structure, a new algorithmUn-suprevised Graph Embedding Structured based on Similarity (Un-SGES) isproposed. The K-means cluster algorithm is used in Un-SGES to establish theembedding graph. Several simulation experiments are taken to illustrate thatUn-SGES preserves the global space structure while feature mapping Theseexperiments also illusteate the effectiveness of Un-SGES in the area of highdimension and multiple elements dataset visualization.Secondly, as for the un-supervied LPP dose not consider the label informationof dataset and is hard to choose the optimal parameter, a new algorithm SupervisedGraph Embedding Structured method based on Correlation (SGEC) is proposed. Thecorrelation and class labels of sample dataset are fully used in SGEC to establish theembedding graph. Several simulation experiments are taken to illustrate that SGECpreserves the class structure while feature mapping.Thirdly, as for LPP preserves the locality geometry structure well while can notpreserve the class structure well, on the contrary, the SGEC preserves the classstructure well while can not preserve the geometry well, a new algorithm LocalStructure Preserving Global Supervised Graph Embedding (LPGSGE) is proposed.This is a fusion algorithm of the former two algorithms. Several similationexperiments are taken to illustrate that LPGSGE preserves both the geometry andclass structures of dataset well while feature mapping.In the end, the proposed algorithms in this thesis are used on the visualizationanalysis of satellite remote sensing data. The time dimension is reduced and thestructure of the sensing parameters are visually presented. The unknown data mode and abnormal state of the satellite has been visually detected. These cases illustratethat the proposed graph embedding structured algorithms can be used in thevisualization of high dimension and multiple elements dataset well.